acceptodds
Under review as a conference paper at ICLR 2027

HDRNet: Highlight-Corrected Dual-Domain Ratio Network for Low-Light Scene Understanding

Abstract

Low-light vision remains a fundamental challenge, as illumination degradation impairs downstream tasks. Recent methods build illumination-invariant color ratios in the log domain under a Lambertian model, and Phong-inspired extensions suppress highlight residuals after ratio formation. However, local highlights introduce channel-dependent biases into the log image before ratio construction. Such biases survive cross-channel subtraction and corrupt the ratio maps in the object regions. We propose HDRNet, a compact plug-and-play highlight-corrected dual-domain ratio network for low-light scene understanding. Local Highlight Attenuation module corrects for highlight-induced log-domain bias by predicting a bounded, channel-wise attenuation map from RGB cues and subtracting it before ratio construction, reducing contamination in the resulting ratio features. The corrected signal feeds a dual-branch module for spatial and spectral feature extraction, and a residual gated fusion integrates both branches with RGB appearance cues. Highlight Prior Regularization supervises the front-end during training. HDRNet delivers mAP gains of 2.2 points on ExDark and 2.6 points on DarkFace, and achieves 42.0% mask AP on LIS and 62.1% mIoU on ACDC-Night. Controlled ratio-recovery experiments support the correction mechanism, while highlight-stratified evaluation shows the largest detection gains under strong background highlights. Code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Low-Light Scene Understanding, Invariant Feature Extraction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_igh1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies an overlooked failure mode: local highlights inject channel-dependent biases into the log image that survive cross-channel subtraction. HDRNet addresses this with LHA (RGB-guided bounded attenuation before ratio construction), dual-domain SLRE/SSR branches (SSR using a real non-negative response for phase preservation), residual-gated fusion, and training-time HPR.

Critical strengths
  1. Clear and useful failure-mode characterization:this paper shows that highlight-induced bias is injected before ratio construction and survives cross-channel subtraction (Eqs. 3–5). This is a sharper observation than the "ratio-then-correct" paradigm of YOLA/FRBNet, and it directly motivates the pre-ratio correction design.
  2. Correction placement is empirically justified: Fig. 5 uses equal-capacity correction heads to isolate the information-availability effect: the RGB-conditioned head reaches AUROC 0.788, whereas the ratio-domain head reaches 0.414 (below chance). The subsequent robustness comparison (mAP drop 0.22 vs 0.45/0.43) makes the argument self-consistent.
  3. Phase-preservation claim and efficiency profile are notable: Constraining FRBNet's sign-unconstrained complex modulation to a real non-negative response is a principled fix, triangulated by controlled phase intervention and filter transplant (Fig. 8). The efficiency result (10.5K params, 12.4 ms, +3.6% FLOPs) is genuinely attractive.
Critical weaknesses
  1. Novelty is largely incremental: SLRE inherits YOLA's zero-mean local filters; SSR is essentially FRBNet's complex multiplier restricted to a real magnitude response; LHA is a learned highlight mask with bounded log-domain subtraction. The combination is reasonable, but the work offers no substantially new modeling insight.
  2. Gains are modest and lack statistical validation: Main tables (1–3) are single-run, with no seeds, variance, or confidence intervals. Ablation deltas (Table 4) are only 0.6–1.9 points and likely fall within run-to-run noise, weakening the "consistent gains" claim. The theoretical analysis (Eqs. 3–5) also gives no magnitude condition or bound for the residual ∆i − ∆j.
Questions for authors

See Weaknesses

Reviewer_Zx8nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HDRNet, a plug-and-play front-end for low-light scene understanding that addresses highlight-induced biases in log-ratio features. It introduces Local Highlight Attenuation (LHA) to correct these biases before ratio construction, followed by Spatial Log-Ratio Extraction (SLRE) and Spectral Suppression and Recalibration (SSR) to capture complementary spatial and frequency-domain features. Highlight Prior Regularization (HPR) further guides selective correction using synthetic highlights. Experiments across multiple low-light recognition tasks demonstrate consistent improvements over existing methods.

Critical strengths
  1. Clear motivation. The authors derive the limitations of log-ratio methods for low-light scene understanding. Local highlight components interfere with the extraction of illumination-invariant features in the log domain, and highlight residuals cannot be eliminated by cross-channel subtraction. This provides a theoretical basis for correcting log-domain signals before constructing ratios.

  2. Effective highlight prior regularization. HPR introduces a physics-driven loss through synthesized highlight perturbations, guiding the network to perform accurate highlight correction without real annotations.

Critical weaknesses
  1. The diversity of the synthetic highlight mode is limited. HPR generates synthetic highlights using a single spatial local Gaussian function. However, in real nighttime scenes, there are usually multiple spatially distributed or overlapping highlights, such as rows formed by street lamps. Although the paper evaluated the robustness under different highlight conditions, it is unclear whether training with a single Gaussian perturbation alone is sufficient to fully capture such complex highlight patterns, which may limit the generalization ability of the proposed regularization strategy.

  2. Potential risk in handling real highlights. Original low-light images may already contain real highlights with unknown positions and intensities. However, HPR mainly constrains the incremental attenuation response to newly injected synthetic highlights, rather than directly supervising the correction of existing highlights. Therefore, even if the synthetic alignment loss is minimized, the predicted attenuation for existing highlights may still be inaccurate. It is necessary to clarify how HPR handles existing highlights, especially when synthetic highlight regions overlap with real highlight regions, and whether this affects correction accuracy.

Questions for authors

1)Sufficiency of synthetic highlight patterns. Is the use of a single Gaussian highlight sufficient to represent the complex highlight patterns encountered in real nighttime scenes, particularly those involving multiple spatially distributed or overlapping highlights? Could the authors provide further justification or experimental evidence to support the adequacy of this simplified synthesis strategy?

  1. Correction of real highlights. Since original training images may already contain real highlights, HPR mainly supervises the incremental attenuation caused by synthetic perturbations, rather than correcting these real highlights. How does the framework ensure accurate attenuation of these real highlights? In particular, what happens when synthetic highlights overlap with real highlights? Does this overlap affect the effectiveness of regularization?
Reviewer_qieqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HDRNet, a plug-and-play front-end for low-light recognition. It starts from the observation that, in existing log-domain color-ratio methods (YOLA, FRBNet), channel-dependent biases induced by local highlights survive cross-channel subtraction and corrupt the ratios. Its core module, LHA, predicts a bounded attenuation map from RGB-based highlight priors and subtracts it from the log image before ratio construction. The corrected signal is processed by two branches, SLRE based on zero-mean spatial filters and SSR based on a real, non-negative radial frequency filter, which are combined through gated fusion, and LHA is supervised during training by HPR losses using synthetic highlights. The paper reports consistent improvements over existing front-ends on detection (ExDark, DarkFace), segmentation (ACDC-Night, LIS), classification (CoDaN), and video recognition (ARID).

Critical strengths

S1. Broad evaluation with consistent gains. Across six tasks (detection, face detection, semantic and instance segmentation, classification, and video recognition) and three detectors (YOLOv3, TOOD, VFNet), HDRNet generally outperforms the recent front-ends YOLA and FRBNet.

S2. Plug-and-play architecture. With 10.5K parameters and about 1.8 additional GFLOPs over the detector, it can be attached without modifying the downstream model.

S3. Clear motivation and diverse analyses. The paper makes many attempts to directly verify the proposed mechanism, including a comparison of correction placement before and after ratio construction, controlled highlight-recovery experiments, and highlight-stratified evaluation.

Critical weaknesses

W1. Limited novelty over existing methods. SLRE directly follows the zero-mean ratio filtering and validity masks of YOLA, and the SLRE-only result (Table 4, 72.7) is almost identical to YOLA (72.9). SSR is a variant of FRBNet's radial-basis frequency filter with a real, non-negative constraint, and the tone consistency loss also follows YOLA. The contribution of the new parts, LHA and HPR, is about +0.7 mAP (ExDark) and +0.9 mAP (DarkFace) (Table 4), and isolating the core claim of "correction before ratio construction" requires attaching LHA to YOLA and FRBNet.

W2. The premise that highlights are a major failure mode on real data is not verified. All mechanism analyses (Figs. 5 and 7, Appendix B.3) use the same synthetic chromatic Gaussian highlights as HPR training, effectively evaluating on the training distribution. Table 11, which uses real images, defines its subsets with HDRNet's own prior (Eq. 7); all methods score about 10 mAP higher on the heavy subsets, and injecting synthetic highlights also increases the mAP of all methods (Fig. 10). Moreover, although the abstract claims that highlights corrupt the ratios in object regions, in the target-inclusive partition, where highlights overlap with objects, the margin over the strongest competing method is smaller on the heavy subset (+2.16) than on the light subset (+2.92).

W3. Weak evidence for the phase-preservation claim. That a real, non-negative filter preserves phase is trivial by construction; what matters is whether this is necessary for performance. However, the evidence consists only of interventions injecting artificial phase shifts (Fig. 8) and a proxy metric comparing the front-end's spectral gradients with semantic boundaries (Boundary F of 15.6% vs. 12.1%), so a comparison of task performance (mAP, mIoU) when SSR is replaced with FRBNet's frequency module within HDRNet is needed.

W4. The gains are marginal when practicality is considered. Latency is critical in the main applications of nighttime detection, such as autonomous driving and surveillance, yet HDRNet's latency increases substantially relative to its parameter count (10.5K) and FLOPs (+1.8 GFLOPs, Table 10) because of the FFT-based SSR. In Fig. 3, methods without additional inference-time architecture run at 7.7 ms, whereas HDRNet runs at 12.4 ms, about 1.6× slower, and the authors themselves acknowledge the additional latency of the FFT branch as a limitation (C.1). With this additional time, competing methods could use a larger backbone or a higher input resolution, but no comparison under equal latency is provided. Taking all of this into account, gains of only +0.8 to +2.6 mAP over FRBNet, which has similar latency (Table 1), are insufficient to constitute a practical advantage.

Minor. The name HDRNet overlaps with the well-known HDRNet [1] and is confusing since the method is unrelated to actual HDR processing. In Fig. 5a, the head that takes ratios as input achieves an AUROC of 0.414, below chance, which suggests that the comparison head may not have been trained properly. The values of the HPR loss weights λ_k and λ_cons are not reported.

References

[1] M. Gharbi et al., "Deep Bilateral Learning for Real-Time Image Enhancement," ACM TOG (SIGGRAPH) 2017.

Questions for authors

Please check the weaknesses.

1 comment

Dear Authors,

Thank you for the detailed rebuttal and the additional experiments. The transfer results in Table 6 and the clarification regarding the loss coefficients address some of my concerns. However, my main concern about novelty remains. As Reviewer igh1 also noted, SLRE and SSR closely follow YOLA and FRBNet, respectively. The proposed LHA and HPR components yield modest gains of 0.6–0.9 mAP when added to existing front ends.

My own research has focused primarily on this area, and I still see the proposed method as an incremental advance over prior work, which I do not find sufficient for ICLR. I will therefore maintain my original score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.